Liquid d1

Liquid d1 returns typed binary decisions, class distributions or scores through Liquid AI’s decision API.

Decisions rather than generated prose

The documented /decisions/v1/systemone interface uses typed questions. Noul returns a binary probability, Choice produces a distribution over defined options and Score returns a scoring distribution. The endpoint does not generate ordinary answer tokens. This is useful for an application that needs a constrained result rather than an open-ended chat response.

Modalities and access

The paid d1 route accepts text and image input. The documented d1:free route is text-only, so it should not be presented as a free equivalent of the image-enabled service. Image requests and the number of questions affect usage under the endpoint’s billing rules. The hosted model’s availability does not establish the licence or hardware requirements of a separately named local release.

Tune an application decision

Start with a clear label definition and examples from the real workflow. Include ambiguous inputs and choose thresholds using measured false-positive and false-negative costs. Inspect whether the returned probabilities are calibrated for your data before treating confidence as an automatic action policy. Keep a review path for uncertain cases.

Use the decision schema directly and validate it at the application boundary. A chat-completion wrapper that expects assistant text is the wrong integration shape. Record the question, class options and threshold with each outcome so a decision can be reviewed later. Do not describe zero generated output tokens as zero-cost service usage.

Sources

Publisher documentation